ADHESION LOSS PREDICTION OF A CLIMBING ROBOT THROUGH ...

ADHESION LOSS PREDICTION OF A CLIMBING ROBOTTHROUGH MAGNETIC FIELD ANALYSIS BY ARTIFICIAL

NEURAL NETWORKS

Rodrigo Valerio EspinozaAndre Schneider de OliveiraLucia Valeria Ramos de ArrudaFlavio Neves JuniorFederal University of Technology, Parana (UTFPR), Automation and Advanced Control System Laboratory (LASCA)

Av. Sete de Setembro, 3165, Reboucas, CEP 80230-901, Curitiba, PR, Brazil

[email protected] [email protected] [email protected] [email protected]

Abstract. This paper discusses an improvement for an autonomous robot’s magnetic climbing system necessaryto perform the inspection of metal plates and weld beads in internal/external LPG storage sphere’s surfaces. Thistask typically covers up to 360◦ in the roll-pitch-yaw angles navigation with respect to the Earth‘s surface plan.Also, the storage spheres’ metal surfaces present non-modeled disturbances (such as surface irregularities, rustdust, welding seams...) that can change the adhesion force, requiring an active control to maintain the forcebalance needed for trajectory tracking. Classic approaches utilize distance sensors to detect gaps between the wheeland the surface, providing a low-level feedback to the control system that, generally, does not have an activecompensation of adhesion, thus, disturbing the robot navigation. Unfortunately this method is inappropriatedue to the adhesion system’s nonlinear behavior (adhesion force between wheel and surface decreases radicallywith gaps) and tracking limitations. In order to surmount these problems, an adhesion disturbances recognitionsystem based in artificial neural networks is developed to predict adherence anomalies through magnetic fieldanalysis. This prediction of adherence loss allows the implementation of a new approach to adhesion feedback inrobot navigation controllers, that could be used to overcome adhesion disturbances. The proposed system fulfillsthe inspection tasks by preventing the detachment of one or more wheels and, possibly, the robot’s fall withoutmaking path deviations.

Keywords: Autonomous Robot, Magnetic Adhesion, Magnetic Field Analysis, Neural Network, DisturbanceRecognition System

1. INTRODUCTION

Industry growth and technology improvement demand an increase of robot-based solutions for a great varietyof processes. Several of the researches realized to supplement this increasing necessity are related with climbingrobots (CR), a multi-task mobile robot system topology applicable to complex environments. These robotscan be used in several fields such as industry petrochemical, nuclear and other power plants, which commonlypresents hazardous tasks.

The CRs can be applied to different purposes, resulting in a wide mechanical topology diversity (e.g. Chuet al. (2010) and Caprari et al. (2012)) of which locomotion and adhesion systems are of greatest interest,given that they represent the main interface with the robot’s work environment. These systems are very closelyattached, hence locomotion and adhesion approaches are not chosen independently. These systems’ interactionresults some specific characteristics which can be crucial in some tasks performance, and also for the projectscope fulfillment.

Common locomotion strategies for climbing tasks are based in legs (e.g., Kim et al. (2007)), tracks (e.g.,Ben-Tzvi et al. (2009)), body translation (e.g., Osswald and Iida (2011)) and wheels (e.g., Schmidt et al. (2011)).The adhesion’s approaches are diverse (e.g, Silva et al. (2008) and Caprari et al. (2012)) but the main strategiesare based in: suction (e.g., Jiang et al. (2009)), gripping (e.g., Guan et al. (2011)), adhesives (e.g., Kim et al.(2007)) and magnetic force (e.g., Xu and Ma (2002)). A classic approach for robot climbing in metallic surfacestasks combines magnetic adhesion with wheel locomotion, e.g., works of Fischer et al. (2011) and Wu et al.(2011).

The inherent characteristic of the magnetic components, typically permanent magnets, guarantees full timeadherence without a power source or other auxiliary device which is very convenient since the CR’s weight is acrucial point in the project. Although this kind of robot is limited to work exclusively on magnetic surfaces, itcan be applied to a broad range of tasks, such as inspection and maintenance tasks in power plants, liquefiedpetroleum gas (LPG) storage spheres and oil tanks, pipelines and ship hulls.

ABCM Symposium Series in Mechatronics - Vol. 6 Copyright © 2014 by ABCM

Part I - International Congress Section V - Mobile Robotics

693

In scientific community, a great diversity of mechanical solutions for climbing robots is discussed, but theapplications are focused on magnetic passive adhesion systems with open loop strategies (e.g. in Fernandezet al. (2010) and Kalra et al. (2006)) which are susceptible to error due to disturbances and can “overthrow”the robot.

The metal surfaces commonly present unmodelled disturbances (such as surface irregularities, rust dust,welding seams, ...) that can change the adhesion force and it requires an active control to maintain the forcebalance needed for trajectory tracking. An alternative approach to overcome this problem is based on discretecontrol of the adhesion force through distance variation between permanent magnets and surface, as seen inWu et al. (2011) and Oliveira et al. (2010). However, this kind of approach is unable to make fast and preciseadjustment of the adhesion (what is often required) due to the nonlinear behavior of the magnets, i.e., a littleincrease in the gap between magnet and surface strongly decreases the adhesion force. Additionally, gap betweenwheel and surface is not the only way to reduce the magnetic adhesion force, other factors such as surface’smaterial magnetic characteristics and surface thickness also reduces the magnetic adhesion force.

This paper presents the active adhesion force control with detachment prediction for CR autonomous navi-gation in a LPG storage sphere. This work is organized in the following manner. Section (2) shows the robot’sconcept. Section (3) describes the robot’s control architecture and Sections (4) and (5) present respectivelyexperimental results and conclusion.

2. ROBOT CONCEPT

2.1 The Inspection Environment and Design Requisites

Storage spheres, as presented in Fig. (1), are metal structures commonly used in the petrochemical sector tostore liquefied gases in high pressures. Their advantage is the reduced dimension compared to cylindrical tankswith the same load capacity, hence cheaper.

Figure 1. LPG storage sphere.

The inspection of LPG tanks is a hazardous and complex process due to the insalubrious environment(especially inside the tank), size of the inspection area and mainly to the height (about 18 m) of the spherestherefore, the process automation is highly recommended.

Autonomous robotic inspection consists in an autonomous navigation task in which a robot covers the entiresurface of the object being inspected and allows the detection of environment’s inconsistencies. In this case, theclimbing robot must navigate in planes including the perpendicular to the Earth’s surface thus, the gravitationaldisturbance is highly important and cannot be neglected.

According to Chu et al. (2010), a climbing robot project must take into account the following aspects:

• Weight : moving against gravity has direct influence of weight force and also the heavier the robot, thebigger is its energy consumption;

• Mobility : in some situations, the robot must be able to move in complicated paths in order to performcertain tasks;

• Adherence: the climbing ability is directly related with the adherence system and its flexibility to dealwith external factors, such as gravity.



694

2.2 Robot Design

The robot’s mechanical structure is showed in Fig. (2), it consists of two parallel sets of fixed wheels (notsteerable). Each set is linked by a Vee belt, hence two motors control the movement of four wheels; the robothas two degrees of freedom. The wheel sets are misaligned so that, when passing small obstacles (e.g., weldseams), two parallel wheels do not decrease their adhesion forces at the same time.

Figure 2. Robot’s mechanical structure.

The wheels were designed so the robot could support its own weight, the inspection equipment and theumbilical cord. Each wheel consists in a set of two ring shaped neodymium magnets positioned between twosteel disks and attached by screws with low magnetic permeability, the set is covered by a high hardnesspolyurethane rubber. The Fig. (3) shows one of the wheels with the transmission pulley attached, its magneticforce is approximately 440N. More details about the robot’s mechanical construction can be found in Rovani(2013) and de Oliveira et al. (2012).

Figure 3. Magnetic wheel’s exploded vision (left) and built magnetic wheel (right).

Figure (4) shows the robot’s devices diagram. The scheme shows that the robot does not have the inspectionmodule. The ultrasonic inspection is an independent module that is only carried by the robot. The mainrequisite of the robot is to perform a smooth navigation that fulfills the navigation requirements allowing astable inspection.



695

Traction controller and Motor power unit

Magnetic wheels

EC motor

Incremental encoder

Planetary gearbox

3-axis Magnetometer

3-axis Accelerometer

3-axis Gyroscope

Inertial navigation system

Exteroceptive sensors

USB

USB

Left Motion Unit

Umbilicalcable

Digital Compass

Outdoor WiFi

Gigabit Ethernet

Hub

AC Power Supply

CAN BUSTraction controller and Motor power unit

Magnetic wheels

EC motor

Incremental encoder

Planetary gearbox

Right Motion Unit

Gigabit Ethernet Interface

Dual DC/DC Converter

MultitreadLinux CPU

USBElectromagnet

Adhesion force controller

Figure 4. Robot’s device diagram.

2.3 Kinematic Model

The robot’s kinematics analysis is made assuming that the robot is situated on a global reference planerepresented by the system (XG, YG, ZG) as exhibited in Fig. (5).

YG

XG

XR

YR

1

2

3

4

θ

a

b

c

ab

c

φ1

φ2

l1

l2

φ1

φ2

l1

l2

Figure 5. Parameters utilized in the robot’s kinematic model.

The robot’s reference system is represented by (XR, YR, ZR)”

positioned in the chassis center and rotatedby de θ in respect of the global system. The robot’s pose is represented in respect to the global system as,

ξG = [ x y θ ]T . (1)

The robot in both systems, ξG and ξR, has same angular velocity (θG = θR) and (x,y) linear velocities relatedby,

ξR = R(θ)ξG, (2)

where R is the rotational matrix,

R(θ) =

cos(θ) sin(θ) 0−sin(θ) cos(θ) 0

0 0 1

. (3)

The robot’s kinematic model can be obtained by the union of kinematic constraints (rolling and sliding) of itswheels. To develop the model, its necessary to describe the wheels with respect to the robot’s system. Table (1)



696

gathers the three parameters that describe the wheels. The parameters are expressed in polar coordinates suchthat l is the distance of the wheels center to the system’s origin O(XR, YR, ZR), α is the rotation angle in ZR

and β is the angle between the wheel plane and the robot’s chassis.

Wheel l α β1 l1 −ϕ1 π/2 + ϕ1

2 l2 −π + ϕ2 3π/2 − ϕ2

3 l2 ϕ2 π/2 − ϕ2

4 l1 π − ϕ1 −π/2 + ϕ1

Table 1. Wheels’ parameters.

The rolling constraint ensure that all wheel rotation results is converted in robot’s displacement (there is noskid), hence,

[Da Db Dc]R(θ)ξG − rϕ = 0, (4)

where, Da

Db

Dc

=

sen(α+ β)−cos(α+ β)−lcos(β)

. (5)

The second constraint imposed is the lateral sliding one, that establishes that the wheel cannot displace inthe plane orthogonal to its,

[Ca Cb Cc]R(θ)ξG = 0, (6)

where, Ca

Cb

Cc

=

cos(α+ β)sen(α+ β)lsen(β)

. (7)

The union of the rolling and sliding constraints (Eqs. (4) and (6)), for every wheel, results the kinematicarrangement,[

DC

]R(θ)ξG =

[JRϕZ

], (8)

where D and C are 4×3 matrices containing the rolling and sliding constraints coefficients , JR is a 4×4 matrixwhich diagonal relates the wheel’s radius, ϕ is a 4× 1 matrix with the wheel’s angular velocities [ϕ1 ϕ1 ϕ2 ϕ2]T

and Z is a 4 × 1 matrix of zeros.The robot’s inverse differential kinematics is defined as,

ϕ = J−1ξG, (9)

where,

J−1 =

[J−1R DC

]R(θ). (10)

From the Eq. (9) the robot’s direct differential kinematics equation is defined,

ξG = Jϕ. (11)

The kinematic model validation is accomplished by a curved trajectory application. The actuator responsibleto move wheels 1 and 2 performs a trapezoidal velocity profile, while the actuator that commands the remainingwheels performs a triangular velocity profile. Figures (6) and (7) shows, respectively, the actuators’ velocityprofiles and the robot’s performed trajectory with origin in (0,0).



697

0 50 100 150 200 250 300 3500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Time [s]

An

gu

lar

velo

city

[ra

d/s

]

Motor 1Motor 2

Figure 6. Motors’ velocity profile.

0 100 200 300

0

50

100

150

200

250

300

350

400

X − Distance [cm]

Y −

Dis

tanc

e [c

m]

Figure 7. Simulated trajectory.

3. CONTROL ARCHITECTURE

The functional behavior of the robot controller is divided in two modes: autonomous and non-autonomous.In non-autonomous mode, the robot is fully controlled by a operator. In autonomous mode, the trajectoryplanning and system tuning are the initial procedures that are processed in a sequential manner. The operationalstages (i.e., navigation and inspection) are executed of parallel manner with processing distributed in dedicatedprocessors. Figure 8 shows the internal behavior of the robot controller.

The climbing strategy is compound by passive and active response, the robot’s wheels act with constantmagnetic characteristics while the electromagnet gives a continuous control over the applied magnetic force.The magnetic force must be controlled when the robot finds and obstacle, such as a welding seam, that makesone of the wheels to loose part of its adherence while overpassing it. Figure (9) shows the adhesion controlsystem, which is detailed in Sections (3.1) and (3.2).



698

Environment identification

System tuning

Trajectory planning

Inspection analysis

Autonomous mode

HMI

Position Mode

Velocity Mode

Start-up

Navigation

Trajectory tracking

Gravitational compensation

Adhesion control

Ultrasonic Inspection

TOFD scanning

Phased array scanning

Non-autonomous mode

Figure 8. Internal behavior of the robot controller.

Pre-processingInertial

Navigation

Sensors

Adhesion

Loss Prediction

Electromagnet

Driver

Adhesion

Compensation

Figure 9. Adhesion control scheme.

3.1 Adhesion Control

The superficial adherence when navigating inside a gas storage sphere depends directly of the gravitationalinfluence on the normal force (FN ) caused by the contact surface. However, mobile robots usually have degreesof freedom parallel to the contact surface and the normal force can restrict the adhesion. To avoid this effects,magnets are often used always in contact with the surface in order to cancel gravitational force (GZ′). Butdisturbances during navigation may temporarily or permanently reduce magnetic adherence, e.g., one of thegoals of gas storage spheres inspection is the welding seam inspection and when passing over these lines thesuperficial adherence is modified. Passive adhesion, i.e., through permanent magnets, do not allow active controlof superficial adhesion and can make the robot fall when a disturbance is occurring. One way of preventingsuch disturbances is the inclusion of an adhesion force active control strategy by the use of passive permanentmagnets (FP ) for usual support and an electromagnet (FA) for the active control of adhesion disturbances, asshown in Fig.(10)

The superficial adherence is designed through Newton’s laws of motion, and its limit is the balance of forcesperpendicular to the surface’s plan acting on rigid body (

∑Fz = 0). However, in balance, any disturbance force

would drop the robot. In order to avoid these effects, a minimum adhesion limit for the system is established,i.e., the reference of the controller will be the minimum adhesion limit (

∑Fz = Fdesired). The equation of

forces for superficial adhesion on the inside of a gas storage sphere is described as

FAdh(t+1) = FN(t)cos(φ(t)) − FP (t) − FA(t) + Fdist(t), (12)

where, FN is the measured normal force, FP is the passive force of permanent magnets, FA is the active forceof electromagnet, Fdist is the force of disturbance and φ is the measured spatial orientation vector.



699

Figure 10. Force balance of active adhesion control.

The electromagnet will be used by the control strategy to maintain the adhesion force greater than theminimum limit (Fdes), i.e., the active adhesion control is modeled as

FA(t+1) = FN(t)cos(φ(t)) − FP (t) + Fdes(t) +KAef , (13)

where, Fdes is the limit of the adhesion force, KA is the adhesion gain and ef is the adhesion error.

3.2 Adhesion Loss Prediction

The adhesion compensation system, in order to work, must receive as a feedback an adhesion error signal.The adhesion loss prediction system objective is to identify the moments when the electromagnet activation isneed therefore, to provide the error signal ef , as expected in Eq. (13).

The difficulty to create a model that solve the prediction problem is significant, because the measuredmagnetic field may radically change due to numerous factors. Depending on the position of the magnetic fieldsensor (installed place in the robot’s chassis), the action of some wheels may be more (closer to the sensor)or less perceived. Ideally the sensor should be closest as possible and equidistant to the 4 wheels, which isnot always possible. Besides the sensor’s position, changes in the robot’s mechanical/structural topology andintroduction of new components may change the profile of the magnetic field measured. These factors encouragethe use of artificial neural networks (ANN) to solve the identification problem.

The objective is to construct a ANN capable to identify magnetic field force profiles enabling the correlationbetween magnetic field force and magnetic adhesion. The target behaviour is showed in Fig. (11).

OFF

ON

Surface with disturbance

Surface without disturbance Magnetic field pattern

Magnetic field pattern

Electromagnet disabled

Electromagnet activated

Normal adhesion situation

Adhesion`s disturbance situation

Figure 11. Adhesion control action.

The designed ANN is a multilayer perceptron with one hidden layer. Figure (12) shows the ANN, where



700

squares represent data and circles represent perceptron neurons. The quantity of neurons and hidden layerswere chosen by performance simulations. The network receives 9 inputs (explained in Tab. (2)), which arefiltered data from the inertial navigation sensor. The output indicates adhesion loss and the wheel in which itis occurring.

Inputs

Hidden

Layer

Output

Layer Outputs

.

.

.

1

2

40

1

2

3

4

1

2

Wheel 1

Wheel 2

3

4

Wheel 3

Wheel 4

1

2

9

Mx

Az

My

3

4

Mxy

Mt

5

6

dMx

dMy

7

8

dMxy

dMt

Figure 12. Designed ANN.

Abbreviation MeaningMX Magnetic field force on the X axisMY Magnetic field force on the Y axisMXY Resultant magnetic field force of the X and Y axisMT Total resultant magnetic field forcedMX Derivative of MX

dMY Derivative of MY

dMXY Derivative of MXY

dMT Derivative of MT

AZ Acceleration on the Z axisTable 2. ANN inputs.

The ANN was trained using the Resilient Backpropagation method, which was the most efficient (fast trainingand lowest error achievement). The training set was composed by 192105 data acquired from several detachmenttests. The error obtained in the best epoch (200) of the training process was 0,63%.



701

4. EXPERIMENTAL RESULTS

The designed ANN was evaluated using the data of 3 new detachment tests. The first 2 tests involvedindependent detachment, i.e. just one wheel perturbed at a time. In the last test more than one will adherencewere reduced at a time. Figures (13) and (14) show, respectively, the ANN experiment for the right (wheels 1and 2) and left wheels set (wheels 3 and 4).

0 100 200 300 400 500 600 700 8000

0.2

0.4

0.6

0.8

1

1.2

Test data

Ad

hes

ion

loss

occ

ure

nce

Target

Wheel 1Wheel 2

0 100 200 300 400 500 600 700 8000

0.2

0.4

0.6

0.8

1

1.2

Test data

Ad

hes

ion

loss

occ

ure

nce

RNA output

Wheel 1Wheel 2

Figure 13. ANN test for wheels 1 and 2.

0 100 200 300 400 500 600 700 8000

0.2

0.4

0.6

0.8

1

1.2

Test data

Ad

hes

ion

loss

occ

ure

nce

Target

Wheel 3Wheel 4

0 100 200 300 400 500 600 700 8000

0.2

0.4

0.6

0.8

1

1.2

Test data

Ad

hes

ion

loss

occ

ure

nce

RNA output

Wheel 3Wheel 4

Figure 14. ANN test for wheels 3 and 4.

The ANN’s efficiency is measured by the similarity between the target and ANN’s output. It can be seenthat the ANN’s outputs profile are very close to the target. Table (3) shows the efficiency of the network.



702

Performance [ % ] Independent disturbances Simultaneous disturbancesWheel 1 97,79 96,14Wheel 2 100 72,39Wheel 3 92,80 77,25Wheel 4 93,93 48,83Overall 98,25 84,34

Table 3. Adhesion loss identification outcome.

5. CONCLUSION

This paper presented a climbing robot intended to be used in inspection of liquefied petroleum gas storagespheres. The system is stabilized by an active adhesion controller which uses a electromagnet as additionalsupport to the passive permanent magnets of the wheels. An artificial neural network is employed to predictadhesion losses by magnetic field analysis, providing a feedback signal to control strategy.

The ANN, applied to adhesion loss prediction, resulted a reliable system capable to fulfill the adhesion controlsystem requisites. Although the system response is binary identifying the detachment occurrence (1 - adhesionloss, 0 - adhesion normal), the experiments results showed that the strategy is robust and ,therefore, is capablefor improvement. The ANN’s system next step will be the prediction of detachment as well as identification ofthe adhesion reduction’s intensity.

Future works will focus in the architecture’s control, intelligent autonomous navigation, sensor fusion andnavigation/mapping system, surface inspection techniques and identification of other parameters by magneticfield analysis.

6. ACKNOWLEDGEMENTS

This project was partially funded by Brazil’s National Counsel of Technological and Scientific Development(CNPq).

7. REFERENCES

Ben-Tzvi, P., Ito, S. and Goldenberg, A., 2009. “A mobile robot with autonomous climbing and descending ofstairs”. Robotica, Cambridge Univ Press, Vol. 27, pp. 171–188.

Caprari, G., Breitenmoser, A., Fischer, W., Hurzeler, C., Tache, F., Siegwart, R., Nguyen, O., Moser, R.,Schoeneich, P. and Mondada, F., 2012. “Highly compact robots for inspection of power plants”. Journal ofField Robotics, Vol. 29, pp. 47–68.

Chu, B., Jung, K., Han, C. and Hong, D., 2010. “A survey of climbing robots: Locomotion and adhesion”.International Journal of Precision Engineering and Manufacturing, Vol. 11, pp. 633–647.

de Oliveira, A.S., de Arruda, L.V.R., Junior, F.N., Espinoza, R.V. and ao Pedro Battistella Nadas, J., 2012.“Adhesion force control and active gravitational compensation for autonomous inspection in lpg storagespheres.” In Robotics Symposium and Latin American Robotics Symposium (SBR-LARS). pp. 232–238.

Fernandez, R., Gonzalez, E., Feliu, V. and Rodrıguez, A., 2010. “A wall climbing robot for tank inspection. anautonomous prototype”. Conference on IEEE Industrial Electronics Society (IECON), pp. 1424–1429.

Fischer, W., Caprari, G., Siegwart, R. and Moser, R., 2011. “Locomotion system for a mobile robot on magneticwheels with both axial and circumferential mobility and with only an 8-mm height for generator inspectionwith the rotor still installed”. IEEE Transactions on Industrial Electronics, Vol. 58, pp. 5296–5303.

Guan, Y., Jiang, L., Zhu, H., Zhou, X., Cai, C., Wu, W., Li, Z., Zhang, H. and Zhang, X., 2011. “Climbot:A modular bio-inspired biped climbing robot”. In Intelligent Robots and Systems (IROS), 2011 IEEE/RSJInternational Conference on. pp. 1473–1478.

Jiang, Z., Li, J., Gao, X., Fan, N. and Wei, B., 2009. “Study on pneumatic wall climbing robot adhesion principleand suction control”. In Robotics and Biomimetics, 2008. ROBIO 2008. IEEE International Conference on.pp. 1812–1817.

Kalra, L., Gu, J. and Meng, M., 2006. “A wall climbing robot for oil tank inspection”. IEEE InternationalConference on Robotics and Biomimetics (ROBIO), pp. 1523–1528.

Kim, S., Spenko, M., Trujillo, S., Heyneman, B., Mattoli, V. and Cutkosky, M.R., 2007. “Whole body adhesion:hierarchical, directional and distributed control of adhesive forces for a climbing robot”. IEEE InternationalConference on Robotics and Automation, pp. 1268 –1273.

Oliveira, A., Silva, M. and Barbosa, R., 2010. “Architecture of an wheeled climbing robot with dynamicadjustment of the adhesion system”. IEEE International Symposium on Intelligent Systems and Informatics



703

(SISY), pp. 127–132.Osswald, M. and Iida, F., 2011. “A climbing robot based on hot melt adhesion”. In Intelligent Robots and

Systems (IROS), 2011 IEEE/RSJ International Conference on. pp. 5107–5112.Rovani, A., 2013. “Desenvolvimento do prototipo de um robo para inspecao de cordoes de solda em superfıcies

metalicas verticais”. Trabalho de Conclusao de Curso - Engenharia Industrial Mecanica - UniversidadeTecnologica Federal do Parana, Curitiba, p. 117.

Schmidt, D., Hillenbrand, C. and Berns, K., 2011. “Omnidirectional locomotion and traction control of thewheel-driven, wall-climbing robot, cromsci”. Robotica, Cambridge Univ Press, Vol. 29, pp. 991–1003.

Silva, M., Machado, J. and Tar, J., 2008. “A survey of technologies for climbing robots adhesion to surfaces”.In Computational Cybernetics, 2008. ICCC 2008. IEEE International Conference on. pp. 127–132.

Wu, M., Gao, X., Yan, W., Fu, Z., Zhao, Y. and Chen, S., 2011. “New mechanism to pass obstacles for magneticclimbing robots with high payload, using only one motor for force-changing and wheel-lifting”. IndustrialRobot: An International Journal, Vol. 38, pp. 372–380.

Xu, Z. and Ma, P., 2002. “A wall-climbing robot for labelling scale of oil tank’s volume”. Robotica, CambridgeUniv Press, Vol. 20, pp. 209–212.

8. RESPONSIBILITY NOTICE

The authors are the only responsible for the printed material included in this paper.



704

ADHESION LOSS PREDICTION OF A CLIMBING ROBOT THROUGH ...

Documents

Transcript of ADHESION LOSS PREDICTION OF A CLIMBING ROBOT THROUGH ...